Merge pull request 'fix(avatar): index knowledge documents asynchronously' (#15) from codex/avatar-knowledge-async-20260904 into main

Reviewed-on: #15

digital-avatar-app/backend/database.py

@@ -53,6 +53,7 @@ def init_db():
         ("knowledge_docs", "embedding_model", "VARCHAR DEFAULT ''"),
         ("knowledge_docs", "chunk_count", "INTEGER DEFAULT 0"),
         ("knowledge_docs", "vectorized_at", "TIMESTAMP"),
+        ("knowledge_docs", "error_message", "VARCHAR DEFAULT ''"),
         ("avatars", "owner_id", "VARCHAR DEFAULT ''"),
         ("authorizations", "takeover_enabled", "BOOLEAN DEFAULT 0"),
         ("authorizations", "takeover_mode", "VARCHAR DEFAULT 'immediate'"),

digital-avatar-app/backend/main.py

@@ -20,6 +20,7 @@ import routers.chat
 import routers.takeover
 from responses import ok
 from services.chat_attachment_service import purge_expired_chat_attachments
+from services.knowledge_vectorizer import knowledge_vectorizer
 from services.token_billing import DEFAULT_TOKEN_GRANT, release_stale_reservations
 
 logger = logging.getLogger(__name__)
@@ -131,6 +132,7 @@ def on_startup():
 
     init_db()
     seed()
+    knowledge_vectorizer.start()
 
     # Release stale resources when startup is invoked again by a reload/test.
     stop_takeover_scheduler()

digital-avatar-app/backend/models.py

@@ -190,6 +190,7 @@ class KnowledgeDoc(Base):
     file_size = Column(Integer, default=0)
     file_url = Column(String, default="")
     status = Column(String, default="uploaded")  # uploaded | parsing | ready | failed
+    error_message = Column(String, default="")  # 建立索引失败原因
     vectorized = Column(Boolean, default=False)  # 是否已向量化
     embedding_model = Column(String, default="")  # 向量模型标识
     chunk_count = Column(Integer, default=0)  # 切片数量
@@ -205,6 +206,7 @@ class KnowledgeDoc(Base):
             "fileSize": self.file_size,
             "fileUrl": self.file_url,
             "status": self.status,
+            "errorMessage": self.error_message or "",
             "vectorized": bool(self.vectorized),
             "embeddingModel": self.embedding_model,
             "chunkCount": self.chunk_count,

digital-avatar-app/backend/routers/knowledge.py

@@ -1,8 +1,5 @@
 import os
-import json
-import logging
 import uuid
-from datetime import datetime, timezone
 
 from fastapi import APIRouter, UploadFile, File, Depends, Header, HTTPException
 from pydantic import BaseModel
@@ -12,10 +9,9 @@ from database import get_db
 from models import KnowledgeDoc, QAPair, KnowledgeChunk, Avatar, User
 from responses import ok, fail
 import embeddings
+from services.knowledge_vectorizer import knowledge_vectorizer
 
 router = APIRouter()
-logger = logging.getLogger(__name__)
-
 BASE_DIR = os.path.dirname(os.path.abspath(__file__))
 UPLOAD_DIR = os.path.abspath(os.getenv("UPLOAD_DIR", os.path.join(BASE_DIR, "uploads")))
 os.makedirs(UPLOAD_DIR, exist_ok=True)
@@ -71,15 +67,6 @@ def list_docs(avatar_id: str, authorization: str = Header(None), db: Session = D
         .order_by(KnowledgeDoc.created_at.desc())
         .all()
     )
-    # Older synchronous uploads could be interrupted after persisting "parsing".
-    # New uploads are committed only after indexing finishes, so these rows are stale.
-    stale_docs = [doc for doc in docs if doc.status == "parsing"]
-    if stale_docs:
-        for doc in stale_docs:
-            doc.status = "failed"
-            doc.vectorized = False
-            doc.chunk_count = 0
-        db.commit()
     return ok([_doc_payload(d) for d in docs])
 
 
@@ -108,50 +95,42 @@ async def upload_doc(avatar_id: str, file: UploadFile = File(...), authorization
         status="parsing",
     )
 
-    # Complete extraction and embedding before the first database commit so a
+    # Persist and acknowledge the upload first. Extraction and embeddings may take
-    # process restart cannot leave a permanent "parsing" row behind.
+    # minutes for a PDF and must never consume the browser request timeout.
-    try:
+    db.add(doc)
-        text = embeddings.extract_text(path, ext)
+    db.commit()
-        chunks = embeddings.chunk_text(text)
+    db.refresh(doc)
-        if not chunks:
+    knowledge_vectorizer.enqueue(doc.id)
-            raise ValueError("文档没有可建立索引的文字内容")
-        vectors = embeddings.embed(chunks)
-        if len(vectors) != len(chunks):
-            raise ValueError("向量服务返回数量与文档分段不一致")
-        doc.vectorized = True
-        doc.embedding_model = embeddings.MODEL
-        doc.chunk_count = len(chunks)
-        doc.vectorized_at = datetime.now(timezone.utc)
-        doc.status = "ready"
-        db.add(doc)
-        for i, (chunk, vector) in enumerate(zip(chunks, vectors)):
-            db.add(
-                KnowledgeChunk(
-                    doc_id=doc.id,
-                    avatar_id=avatar_id,
-                    content=chunk,
-                    vector=json.dumps(vector),
-                    chunk_index=i,
-                    embedding_model=embeddings.MODEL,
-                )
-            )
-        db.commit()
-        db.refresh(doc)
-    except Exception as exc:
-        db.rollback()
-        doc.status = "failed"
-        doc.vectorized = False
-        doc.embedding_model = ""
-        doc.chunk_count = 0
-        doc.vectorized_at = None
-        db.add(doc)
-        db.commit()
-        db.refresh(doc)
-        logger.exception("knowledge vectorization failed for %s: %s", doc.id, exc)
 
     return ok(_doc_payload(doc))
 
 
+@router.post("/avatar/{avatar_id}/knowledge/docs/{doc_id}/retry")
+def retry_doc(avatar_id: str, doc_id: str, authorization: str = Header(None), db: Session = Depends(get_db)):
+    _require_owned_avatar(db, avatar_id, authorization)
+    doc = db.query(KnowledgeDoc).filter(
+        KnowledgeDoc.id == doc_id, KnowledgeDoc.avatar_id == avatar_id
+    ).first()
+    if not doc:
+        return fail("文档不存在", code=404)
+    if doc.vectorized and doc.status == "ready":
+        return ok(_doc_payload(doc))
+    stored_name = os.path.basename(doc.file_url or "")
+    if not stored_name or not os.path.isfile(os.path.join(UPLOAD_DIR, avatar_id, stored_name)):
+        return fail("原文件不可用，请重新上传", code=400)
+    db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == doc.id).delete()
+    doc.status = "parsing"
+    doc.vectorized = False
+    doc.embedding_model = ""
+    doc.chunk_count = 0
+    doc.vectorized_at = None
+    doc.error_message = ""
+    db.commit()
+    db.refresh(doc)
+    knowledge_vectorizer.enqueue(doc.id)
+    return ok(_doc_payload(doc))
+
+
 @router.delete("/avatar/{avatar_id}/knowledge/docs/{doc_id}")
 def delete_doc(avatar_id: str, doc_id: str, authorization: str = Header(None), db: Session = Depends(get_db)):
     _require_owned_avatar(db, avatar_id, authorization)

digital-avatar-app/backend/services/knowledge_vectorizer.py

@@ -0,0 +1,125 @@
+"""Durable, serial knowledge-document indexing for the avatar knowledge base."""
+
+import json
+import logging
+import os
+import queue
+import threading
+from datetime import datetime, timezone
+
+from database import SessionLocal
+from models import KnowledgeChunk, KnowledgeDoc
+import embeddings
+
+logger = logging.getLogger(__name__)
+
+BACKEND_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+UPLOAD_DIR = os.path.abspath(
+    os.getenv("UPLOAD_DIR", os.path.join(BACKEND_DIR, "routers", "uploads"))
+)
+
+
+class KnowledgeVectorizer:
+    """Indexes one document at a time so slow providers cannot block uploads."""
+
+    def __init__(self):
+        self._queue: queue.Queue[str] = queue.Queue()
+        self._queued: set[str] = set()
+        self._lock = threading.Lock()
+        self._thread: threading.Thread | None = None
+
+    def start(self):
+        if self._thread and self._thread.is_alive():
+            return
+        self._thread = threading.Thread(
+            target=self._run, name="knowledge-vectorizer", daemon=True
+        )
+        self._thread.start()
+        db = SessionLocal()
+        try:
+            # A process restart must not abandon documents already accepted by upload.
+            for (doc_id,) in db.query(KnowledgeDoc.id).filter(KnowledgeDoc.status == "parsing"):
+                self.enqueue(doc_id)
+        finally:
+            db.close()
+
+    def enqueue(self, doc_id: str):
+        with self._lock:
+            if doc_id in self._queued:
+                return
+            self._queued.add(doc_id)
+        self._queue.put(doc_id)
+
+    def _run(self):
+        while True:
+            doc_id = self._queue.get()
+            try:
+                self.vectorize_document(doc_id)
+            except Exception:
+                logger.exception("Unexpected knowledge vectorizer failure for %s", doc_id)
+            finally:
+                with self._lock:
+                    self._queued.discard(doc_id)
+                self._queue.task_done()
+
+    def vectorize_document(self, doc_id: str):
+        db = SessionLocal()
+        try:
+            doc = db.get(KnowledgeDoc, doc_id)
+            if not doc or doc.status != "parsing":
+                return
+
+            stored_name = os.path.basename(doc.file_url or "")
+            path = os.path.join(UPLOAD_DIR, doc.avatar_id, stored_name)
+            if not stored_name or not os.path.isfile(path):
+                raise FileNotFoundError("原文件不可用，请重新上传")
+
+            text = embeddings.extract_text(path, f".{doc.file_type}")
+            chunks = embeddings.chunk_text(text)
+            if not chunks:
+                raise ValueError("文档没有可建立索引的文字内容")
+            vectors = embeddings.embed(chunks)
+            if len(vectors) != len(chunks):
+                raise ValueError("向量服务返回数量与文档分段不一致")
+
+            # Commit the document and every chunk together. Chat only sees complete indexes.
+            db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == doc.id).delete()
+            db.add_all(
+                [
+                    KnowledgeChunk(
+                        doc_id=doc.id,
+                        avatar_id=doc.avatar_id,
+                        content=chunk,
+                        vector=json.dumps(vector),
+                        chunk_index=index,
+                        embedding_model=embeddings.MODEL,
+                    )
+                    for index, (chunk, vector) in enumerate(zip(chunks, vectors))
+                ]
+            )
+            doc.vectorized = True
+            doc.embedding_model = embeddings.MODEL
+            doc.chunk_count = len(chunks)
+            doc.vectorized_at = datetime.now(timezone.utc)
+            doc.status = "ready"
+            doc.error_message = ""
+            db.commit()
+            logger.info("Knowledge document %s indexed with %s chunks", doc.id, len(chunks))
+        except Exception as exc:
+            db.rollback()
+            failed_doc = db.get(KnowledgeDoc, doc_id)
+            if failed_doc:
+                db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == failed_doc.id).delete()
+                failed_doc.status = "failed"
+                failed_doc.vectorized = False
+                failed_doc.embedding_model = ""
+                failed_doc.chunk_count = 0
+                failed_doc.vectorized_at = None
+                failed_doc.error_message = str(exc)[:300] or "建立知识索引失败"
+                db.commit()
+            logger.exception("Knowledge vectorization failed for %s: %s", doc_id, exc)
+        finally:
+            db.close()
+
+
+knowledge_vectorizer = KnowledgeVectorizer()

digital-avatar-app/backend/tests/test_knowledge_storage.py

@@ -8,6 +8,7 @@ from database import SessionLocal
 from main import app
 from models import Avatar, KnowledgeChunk, KnowledgeDoc, QAPair
 from routers.knowledge import _doc_payload
+from services.knowledge_vectorizer import knowledge_vectorizer
 
 
 client = TestClient(app)
@@ -31,14 +32,14 @@ def test_doc_payload_reports_whether_the_persisted_file_exists(tmp_path: Path):
         assert _doc_payload(doc)["filePresent"] is True
 
 
-def test_upload_marks_vectorization_failure_instead_of_staying_processing(
+def test_upload_returns_before_background_vectorization(
     tmp_path: Path,
     authorization_context,
 ):
     context = authorization_context
     with (
         patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
-        patch("routers.knowledge.embeddings.embed", side_effect=RuntimeError("provider unavailable")),
+        patch("routers.knowledge.knowledge_vectorizer.enqueue") as enqueue,
     ):
         response = client.post(
             f"/api/avatar/{context['avatar'].id}/knowledge/docs",
@@ -47,14 +48,15 @@ def test_upload_marks_vectorization_failure_instead_of_staying_processing(
         )
 
     payload = response.json()["data"]
-    assert payload["status"] == "failed"
+    assert payload["status"] == "parsing"
     assert payload["vectorized"] is False
     assert payload["chunkCount"] == 0
+    enqueue.assert_called_once_with(payload["id"])
 
     db = SessionLocal()
     try:
         stored = db.query(KnowledgeDoc).filter(KnowledgeDoc.id == payload["id"]).one()
-        assert stored.status == "failed"
+        assert stored.status == "parsing"
         assert db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == stored.id).count() == 0
         db.delete(stored)
         db.commit()
@@ -62,14 +64,14 @@ def test_upload_marks_vectorization_failure_instead_of_staying_processing(
         db.close()
 
 
-def test_markdown_upload_commits_ready_document_and_chunks_together(
+def test_background_vectorizer_commits_ready_document_and_chunks_together(
     tmp_path: Path,
     authorization_context,
 ):
     context = authorization_context
     with (
         patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
-        patch("routers.knowledge.embeddings.embed", return_value=[[1.0, 0.0]]),
+        patch("routers.knowledge.knowledge_vectorizer.enqueue"),
     ):
         response = client.post(
             f"/api/avatar/{context['avatar'].id}/knowledge/docs",
@@ -78,14 +80,19 @@ def test_markdown_upload_commits_ready_document_and_chunks_together(
         )
 
     payload = response.json()["data"]
-    assert payload["status"] == "ready"
+    assert payload["status"] == "parsing"
-    assert payload["vectorized"] is True
+    with (
-    assert payload["chunkCount"] == 1
+        patch("services.knowledge_vectorizer.UPLOAD_DIR", str(tmp_path)),
+        patch("services.knowledge_vectorizer.embeddings.embed", return_value=[[1.0, 0.0]]),
+    ):
+        knowledge_vectorizer.vectorize_document(payload["id"])
 
     db = SessionLocal()
     try:
         stored = db.query(KnowledgeDoc).filter(KnowledgeDoc.id == payload["id"]).one()
         assert stored.status == "ready"
+        assert stored.vectorized is True
+        assert stored.chunk_count == 1
         assert db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == stored.id).count() == 1
         db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == stored.id).delete()
         db.delete(stored)
@@ -94,6 +101,87 @@ def test_markdown_upload_commits_ready_document_and_chunks_together(
         db.close()
 
 
+def test_background_vectorizer_keeps_failure_reason_for_retry(
+    tmp_path: Path,
+    authorization_context,
+):
+    context = authorization_context
+    with (
+        patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
+        patch("routers.knowledge.knowledge_vectorizer.enqueue"),
+    ):
+        response = client.post(
+            f"/api/avatar/{context['avatar'].id}/knowledge/docs",
+            headers=context["owner_headers"],
+            files={"file": ("knowledge.md", b"# Knowledge\n\nTest content", "text/markdown")},
+        )
+
+    payload = response.json()["data"]
+    with (
+        patch("services.knowledge_vectorizer.UPLOAD_DIR", str(tmp_path)),
+        patch("services.knowledge_vectorizer.embeddings.embed", side_effect=RuntimeError("provider unavailable")),
+    ):
+        knowledge_vectorizer.vectorize_document(payload["id"])
+
+    db = SessionLocal()
+    try:
+        stored = db.query(KnowledgeDoc).filter(KnowledgeDoc.id == payload["id"]).one()
+        assert stored.status == "failed"
+        assert stored.error_message == "provider unavailable"
+        assert db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == stored.id).count() == 0
+        db.delete(stored)
+        db.commit()
+    finally:
+        db.close()
+
+
+def test_retry_queues_a_failed_document_again(
+    tmp_path: Path,
+    authorization_context,
+):
+    context = authorization_context
+    document_id = f"retry-doc-{context['suffix']}"
+    avatar_dir = tmp_path / context["avatar"].id
+    avatar_dir.mkdir()
+    (avatar_dir / "retry.md").write_text("retry content", encoding="utf-8")
+    db = SessionLocal()
+    try:
+        db.add(
+            KnowledgeDoc(
+                id=document_id,
+                avatar_id=context["avatar"].id,
+                filename="retry.md",
+                file_type="md",
+                file_url=f"/api/files/{context['avatar'].id}/retry.md",
+                status="failed",
+                error_message="provider unavailable",
+            )
+        )
+        db.commit()
+    finally:
+        db.close()
+
+    with (
+        patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
+        patch("routers.knowledge.knowledge_vectorizer.enqueue") as enqueue,
+    ):
+        response = client.post(
+            f"/api/avatar/{context['avatar'].id}/knowledge/docs/{document_id}/retry",
+            headers=context["owner_headers"],
+        )
+
+    payload = response.json()["data"]
+    assert payload["status"] == "parsing"
+    assert payload["errorMessage"] == ""
+    enqueue.assert_called_once_with(document_id)
+    db = SessionLocal()
+    try:
+        db.query(KnowledgeDoc).filter(KnowledgeDoc.id == document_id).delete()
+        db.commit()
+    finally:
+        db.close()
+
+
 def test_each_avatar_has_an_independent_document_and_qa_scope(authorization_context):
     context = authorization_context
     first_avatar_id = context["avatar"].id

digital-avatar-app/src/api/index.ts

@@ -305,6 +305,7 @@ export interface KnowledgeDoc {
   vectorized?: boolean
   embeddingModel?: string
   chunkCount?: number
+  errorMessage?: string
   createdAt: string
 }
 
@@ -335,7 +336,8 @@ export const uploadKnowledgeDoc = (avatarId: string, file: File) => {
   const form = new FormData()
   form.append('file', file)
   return request.post<KnowledgeDoc>(`/avatar/${avatarId}/knowledge/docs`, form, {
-    headers: { 'Content-Type': 'multipart/form-data' }
+    headers: { 'Content-Type': 'multipart/form-data' },
+    timeout: 120000
   })
 }
 
@@ -343,6 +345,9 @@ export const uploadKnowledgeDoc = (avatarId: string, file: File) => {
 export const deleteKnowledgeDoc = (avatarId: string, docId: string) =>
   request.delete(`/avatar/${avatarId}/knowledge/docs/${docId}`)
 
+export const retryKnowledgeDoc = (avatarId: string, docId: string) =>
+  request.post<KnowledgeDoc>(`/avatar/${avatarId}/knowledge/docs/${docId}/retry`)
+
 // 标准问答对列表
 export const getQAPairs = (avatarId: string) =>
   request.get<QAPair[]>(`/avatar/${avatarId}/knowledge/qa`)

digital-avatar-app/src/views/KnowledgeManage.vue

@@ -27,7 +27,7 @@
             <p class="upload-hint">支持 MD / TXT / PDF / DOC / DOCX / XLSX，上传后自动向量化</p>
             <input ref="fileInput" type="file" accept=".md,.txt,.pdf,.doc,.docx,.xlsx" class="hidden-input" @change="onFileChange" />
           </div>
-          <p v-if="uploading" class="uploading-text">上传并向量化中…</p>
+          <p v-if="uploading" class="uploading-text">文件上传中…</p>
           <p v-if="uploadError" class="error-text">{{ uploadError }}</p>
         </div>
 
@@ -42,7 +42,10 @@
               <p class="card-meta">{{ doc.fileType.toUpperCase() }} · {{ formatSize(doc.fileSize) }} · {{ formatDate(doc.createdAt) }}</p>
               <p class="card-detail">{{ documentState(doc).detail }}</p>
             </div>
-            <button class="card-delete" @click="removeDoc(doc.id)">删除</button>
+            <div class="card-actions">
+              <button v-if="documentState(doc).tone === 'failed'" class="card-retry" @click="retryDoc(doc.id)">重新索引</button>
+              <button class="card-delete" @click="removeDoc(doc.id)">删除</button>
+            </div>
           </article>
         </div>
         <div v-else class="card-empty">📂 暂无文档，先上传一个知识文件</div>
@@ -78,7 +81,7 @@
 </template>
 
 <script setup lang="ts">
-import { ref, onMounted, computed } from 'vue'
+import { ref, onMounted, onUnmounted, computed } from 'vue'
 import { useRoute, useRouter } from 'vue-router'
 import { useAvatarStore } from '@/store/avatar'
 import { pickScopedAvatarId, unwrapListData } from '@/utils/avatar-page-data.js'
@@ -87,6 +90,7 @@ import {
   getKnowledgeDocs,
   uploadKnowledgeDoc,
   deleteKnowledgeDoc,
+  retryKnowledgeDoc,
   getQAPairs,
   deleteQAPair,
   searchKnowledge,
@@ -107,6 +111,7 @@ const uploading = ref(false)
 const uploadError = ref('')
 const dragOver = ref(false)
 const fileInput = ref<HTMLInputElement | null>(null)
+let documentPollingTimer: ReturnType<typeof setInterval> | undefined
 
 const query = ref('')
 const searching = ref(false)
@@ -123,7 +128,26 @@ const documentState = (doc: any) => {
   if (['uploaded', 'parsing'].includes(String(doc.status || '').toLowerCase())) {
     return { tone: 'pending', label: '处理中', detail: '正在解析并建立知识索引' }
   }
-  return { tone: 'failed', label: '处理失败', detail: '未能建立知识索引，请删除后重新上传' }
+  return { tone: 'failed', label: '处理失败', detail: doc.errorMessage || '未能建立知识索引，请重新索引或重新上传' }
+}
+
+const hasPendingDocuments = () => docs.value.some((doc) =>
+  ['uploaded', 'parsing'].includes(String(doc.status || '').toLowerCase())
+)
+
+const stopDocumentPolling = () => {
+  if (documentPollingTimer) {
+    clearInterval(documentPollingTimer)
+    documentPollingTimer = undefined
+  }
+}
+
+const startDocumentPolling = () => {
+  if (documentPollingTimer || !hasPendingDocuments()) return
+  documentPollingTimer = setInterval(async () => {
+    await loadDocs()
+    if (!hasPendingDocuments()) stopDocumentPolling()
+  }, 2000)
 }
 
 const loadDocs = async () => {
@@ -131,6 +155,7 @@ const loadDocs = async () => {
   try {
     const res: any = await getKnowledgeDocs(avatarId.value)
     docs.value = unwrapListData(res)
+    startDocumentPolling()
   } catch (e) {
     console.error(e)
   }
@@ -182,6 +207,17 @@ const doUpload = async (file: File) => {
   }
 }
 
+const retryDoc = async (id: string) => {
+  if (!avatarId.value) return
+  uploadError.value = ''
+  try {
+    await retryKnowledgeDoc(avatarId.value, id)
+    await loadDocs()
+  } catch (e: any) {
+    uploadError.value = e?.message || '重新索引失败'
+  }
+}
+
 const removeDoc = async (id: string) => {
   if (!avatarId.value) return
   await deleteKnowledgeDoc(avatarId.value, id)
@@ -257,6 +293,8 @@ onMounted(async () => {
   if (avatarId.value) store.currentAvatarId = avatarId.value
   await Promise.all([loadDocs(), loadQA()])
 })
+
+onUnmounted(stopDocumentPolling)
 </script>
 
 <style scoped>
@@ -300,7 +338,10 @@ onMounted(async () => {
 .status-pill.missing { color: #B91C1C; background: #FEF2F2; }
 .status-pill.failed { color: #B91C1C; background: #FEF2F2; }
 .card-meta, .card-detail { margin: 5px 0 0; color: #9398AE; font-size: 11px; line-height: 1.4; }.card-detail { color: #8B6B58; }
-.card-delete { flex: 0 0 auto; align-self: center; border: 0; color: #EF4444; background: #FEF2F2; border-radius: 8px; padding: 7px 9px; font-size: 12px; cursor: pointer; }
+.card-actions { flex: 0 0 auto; display: flex; flex-direction: column; align-items: stretch; gap: 6px; }
+.card-delete, .card-retry { align-self: center; border: 0; border-radius: 8px; padding: 7px 9px; font-size: 12px; cursor: pointer; white-space: nowrap; }
+.card-delete { color: #EF4444; background: #FEF2F2; }
+.card-retry { color: #C15F18; background: #FFF3E6; }
 .card-empty { padding: 42px 16px; border: 1px dashed #F1D9C3; border-radius: 16px; color: #9398AE; background: #fff; font-size: 14px; text-align: center; }
 .qa-card { align-items: stretch; text-align: left; }.qa-card.qa-disabled { opacity: .58; }
 .qa-card .card-content,